Pixel Art Portraits
Consider supporting further research on Patreon or Twitter
Checkout my other models at RetroDiffusion
Trained with 64x64 samples, all generations will need to be downscaled 8 times.
512x512 => 64x64 with Nearest Neighbors. You can use an extension or Aseprite.
No extra pixelizations methods required.
Best used in conjunction with: Astropulse PixelDetector
Tips:
Used best with DreamShaper
Use it at 0.9-0.8 of strength
Better results at 512x512, 512x768 and 768x512
Use simple prompts and no trigger word, see examples for prompt examples
Changelog
v1: Release of v1 checkpoint, a bit overffited on man/woman. Can generalize well
Description
Initial release
FAQ
Comments (18)
Thank you for making this with no trigger word, really helps the workflow. Looks good!
You are the best, perfect pixel art lora. Thank you very much
Thank you so much for this! Fantastic work.
beautiful lora!
The best pixel lora I've ever used
honestly, kinda amazing at emulating a certain farming sim game, good job!
how to use this in img2img, can someone share a good workflow?
img2img + ControlNet Tile + 1.0 Denoising Strength gives you the best results :)
Take a look at the Mona Lisa examples I posted on the gallery, they should have generation params
Holy shit, the base lora is great on its own but if you follow this workflow the images are INSANE. I can't believe this is generated pixel art it looks PHENOMENAL.
@NeriJS ControlNet Tile with what in a1111? The Mona Lisa one is confusing!
@wrc35 Yup, you place the original image in img2img, set Denoising Strength to ~0.8 and enable ControlNet reference or tile
I really like your work, it's very nice of you to share so thank you ! :)
Thank you :)
perfect, easy to use
using the lora control extension https://github.com/cheald/sd-webui-loractl you can do stuff like <lora:svportrait64-v1:0@10,1@11> to turn a normal image pixelated after 10 steps, super neat
And if you use the https://github.com/AUTOMATIC1111/stable-diffusion-webui-pixelization extension, the default pixel size of 4 works perfectly to perfect the image.
Example posted in reviews.
edit: completely didn't notice your python pixelizer, will test.
edit 2: ok so i might have to write a bash script to make using the python pixelizer quicker, but it seems really great, i love how you can pick how many colors to use in the palette
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.